INTEGRATING CROP PHENOLOGY INTO PROBABILISTIC DEEP LEARNING FOR RELIABLE YIELD FORECASTS

Predictive analytics for crop yield in a variable climate environment is vital for ensuring food safety and proper management of agricultural risks, but it is a challenge that deterministic modeling typically fails to convey predictive confidence. This work proposes a Phenologyguided Bayesian Convolutional Neural Network (PB CNN) that combines remote sensingbased vegetation indicators, gridded meteorological datasets, and soil data, along with explicit crop developmental stage data to provide not only point estimates but decomposed estimates of uncertainty. The proposed PB CNN model was trained and evaluated using 8,243 soybean yield observations (2003–2018) for US counties from the Corn Belt states, utilizing the yearbased cross-validation procedure which is representative of real-world season-ahead forecasting. Our PB CNN achieves an RMSE of 0.29 t ha⁻¹, R² of 0.80, and MAE of 0.22 t ha⁻¹, outperforming alternative models such as gradient boosting, a deterministic convolutional neural network, and a baseline Bayesian neural network. If we classify soybean yield anomalies into below, near, and above trend categories, then the model provides 82.4% classification accuracy, 0.82 macro-averaged precision, 0.81 recall, and an F1-score of 0.81, while having severe miss-classification rates lower than 2%. It is also worth noting that the 90% prediction intervals obtained by our model have an empirical coverage probability of 0.91, matching the nominal level, and a mean interval width of 1.05 t ha⁻¹. Satellite greenness metrics during reproductive growth stages, blooming to pod setting, are found to be most important predictors via a permutation-based approach, agreeing with physiological reasoning. A lightweight Gradio-based inference platform allows users to get a probabilistic forecast in less than two seconds, making this tool relevant for timely insurance and commodity trading decisions. Thus, we propose a robust and operational forecasting solution for crop yields under climate uncertainty.